int
codel_add_altq(struct ifnet *ifp, struct pf_altq *a)
{
struct codel_if *cif;
struct codel_opts *opts;
if (ifp == NULL)
return (EINVAL);
if (!ALTQ_IS_READY(&ifp->if_snd))
return (ENODEV);
opts = &a->pq_u.codel_opts;
cif = malloc(sizeof(struct codel_if), M_DEVBUF, M_NOWAIT | M_ZERO);
if (cif == NULL)
return (ENOMEM);
cif->cif_bandwidth = a->ifbandwidth;
cif->cif_ifq = &ifp->if_snd;
cif->cl_q = malloc(sizeof(class_queue_t), M_DEVBUF, M_NOWAIT | M_ZERO);
if (cif->cl_q == NULL) {
free(cif, M_DEVBUF);
return (ENOMEM);
}
if (a->qlimit == 0)
a->qlimit = 50; /* use default. */
qlimit(cif->cl_q) = a->qlimit;
qtype(cif->cl_q) = Q_CODEL;
qlen(cif->cl_q) = 0;
qsize(cif->cl_q) = 0;
if (opts->target == 0)
opts->target = 5;
if (opts->interval == 0)
opts->interval = 100;
cif->codel.params.target = machclk_freq * opts->target / 1000;
cif->codel.params.interval = machclk_freq * opts->interval / 1000;
cif->codel.params.ecn = opts->ecn;
cif->codel.stats.maxpacket = 256;
cif->cl_stats.qlength = qlen(cif->cl_q);
cif->cl_stats.qlimit = qlimit(cif->cl_q);
/* keep the state in pf_altq */
a->altq_disc = cif;
return (0);
}
static int
priq_addq(struct priq_class *cl, struct mbuf *m)
{
#ifdef ALTQ_RIO
if (q_is_rio(cl->cl_q))
return rio_addq((rio_t *)cl->cl_red, cl->cl_q, m,
cl->cl_pktattr);
#endif
#ifdef ALTQ_RED
if (q_is_red(cl->cl_q))
return red_addq(cl->cl_red, cl->cl_q, m, cl->cl_pktattr);
#endif
if (qlen(cl->cl_q) >= qlimit(cl->cl_q)) {
m_freem(m);
return (-1);
}
if (cl->cl_flags & PRCF_CLEARDSCP)
write_dsfield(m, cl->cl_pktattr, 0);
_addq(cl->cl_q, m);
return (0);
}
/* copy the stats info in rm_class to class_states_t */
static void
get_class_stats(class_stats_t *statsp, struct rm_class *cl)
{
statsp->xmit_cnt = cl->stats_.xmit_cnt;
statsp->drop_cnt = cl->stats_.drop_cnt;
statsp->over = cl->stats_.over;
statsp->borrows = cl->stats_.borrows;
statsp->overactions = cl->stats_.overactions;
statsp->delays = cl->stats_.delays;
statsp->depth = cl->depth_;
statsp->priority = cl->pri_;
statsp->maxidle = cl->maxidle_;
statsp->minidle = cl->minidle_;
statsp->offtime = cl->offtime_;
statsp->qmax = qlimit(cl->q_);
statsp->ns_per_byte = cl->ns_per_byte_;
statsp->wrr_allot = cl->w_allotment_;
statsp->qcnt = qlen(cl->q_);
statsp->avgidle = cl->avgidle_;
statsp->qtype = qtype(cl->q_);
#ifdef ALTQ_RED
if (q_is_red(cl->q_))
red_getstats(cl->red_, &statsp->red[0]);
#endif
#ifdef ALTQ_RIO
if (q_is_rio(cl->q_))
rio_getstats((rio_t *)cl->red_, &statsp->red[0]);
#endif
}
int
codel_addq(struct codel *c, class_queue_t *q, struct mbuf *m)
{
struct m_tag *mtag;
uint64_t *enqueue_time;
if (qlen(q) < qlimit(q)) {
mtag = m_tag_locate(m, MTAG_CODEL, 0, NULL);
if (mtag == NULL)
mtag = m_tag_alloc(MTAG_CODEL, 0, sizeof(uint64_t),
M_NOWAIT);
if (mtag == NULL) {
m_freem(m);
return (-1);
}
enqueue_time = (uint64_t *)(mtag + 1);
*enqueue_time = read_machclk();
m_tag_prepend(m, mtag);
_addq(q, m);
return (0);
}
c->drop_overlimit++;
m_freem(m);
return (-1);
}
int
rmc_modclass(struct rm_class *cl, u_int nsecPerByte, int maxq, u_int maxidle,
int minidle, u_int offtime, int pktsize)
{
struct rm_ifdat *ifd;
u_int old_allotment;
int s;
ifd = cl->ifdat_;
old_allotment = cl->allotment_;
#ifdef __NetBSD__
s = splnet();
#else
s = splimp();
#endif
IFQ_LOCK(ifd->ifq_);
cl->allotment_ = RM_NS_PER_SEC / nsecPerByte; /* Bytes per sec */
cl->qthresh_ = 0;
cl->ns_per_byte_ = nsecPerByte;
qlimit(cl->q_) = maxq;
#if 1 /* minidle is also scaled in ALTQ */
cl->minidle_ = (minidle * nsecPerByte) / 8;
if (cl->minidle_ > 0)
cl->minidle_ = 0;
#else
cl->minidle_ = minidle;
#endif
cl->maxidle_ = (maxidle * nsecPerByte) / 8;
if (cl->maxidle_ == 0)
cl->maxidle_ = 1;
#if 1 /* offtime is also scaled in ALTQ */
cl->avgidle_ = cl->maxidle_;
cl->offtime_ = ((offtime * nsecPerByte) / 8) >> RM_FILTER_GAIN;
if (cl->offtime_ == 0)
cl->offtime_ = 1;
#else
cl->avgidle_ = 0;
cl->offtime_ = (offtime * nsecPerByte) / 8;
#endif
/*
* If CBQ's WRR is enabled, then initialize the class WRR state.
*/
if (ifd->wrr_) {
ifd->alloc_[cl->pri_] += cl->allotment_ - old_allotment;
rmc_wrr_set_weights(ifd);
}
IFQ_UNLOCK(ifd->ifq_);
splx(s);
return (0);
}
static void
get_class_stats(struct priq_classstats *sp, struct priq_class *cl)
{
sp->class_handle = cl->cl_handle;
sp->qlength = qlen(cl->cl_q);
sp->qlimit = qlimit(cl->cl_q);
sp->period = cl->cl_period;
sp->xmitcnt = cl->cl_xmitcnt;
sp->dropcnt = cl->cl_dropcnt;
sp->qtype = qtype(cl->cl_q);
#ifdef ALTQ_RED
if (q_is_red(cl->cl_q))
red_getstats(cl->cl_red, &sp->red[0]);
#endif
#ifdef ALTQ_RIO
if (q_is_rio(cl->cl_q))
rio_getstats((rio_t *)cl->cl_red, &sp->red[0]);
#endif
}
int
priq_get_class_stats(struct priq_if *pif, u_int32_t qid,
struct priq_classstats *sp)
{
struct priq_class *cl;
IFCQ_LOCK_ASSERT_HELD(pif->pif_ifq);
if ((cl = priq_clh_to_clp(pif, qid)) == NULL)
return (EINVAL);
sp->class_handle = cl->cl_handle;
sp->priority = cl->cl_pri;
sp->qlength = qlen(&cl->cl_q);
sp->qlimit = qlimit(&cl->cl_q);
sp->period = cl->cl_period;
sp->xmitcnt = cl->cl_xmitcnt;
sp->dropcnt = cl->cl_dropcnt;
sp->qtype = qtype(&cl->cl_q);
sp->qstate = qstate(&cl->cl_q);
#if CLASSQ_RED
if (q_is_red(&cl->cl_q))
red_getstats(cl->cl_red, &sp->red[0]);
#endif /* CLASSQ_RED */
#if CLASSQ_RIO
if (q_is_rio(&cl->cl_q))
rio_getstats(cl->cl_rio, &sp->red[0]);
#endif /* CLASSQ_RIO */
#if CLASSQ_BLUE
if (q_is_blue(&cl->cl_q))
blue_getstats(cl->cl_blue, &sp->blue);
#endif /* CLASSQ_BLUE */
if (q_is_sfb(&cl->cl_q) && cl->cl_sfb != NULL)
sfb_getstats(cl->cl_sfb, &sp->sfb);
return (0);
}
static inline int
fairq_addq(struct fairq_class *cl, struct mbuf *m, struct pf_mtag *t)
{
struct ifclassq *ifq = cl->cl_fif->fif_ifq;
fairq_bucket_t *b;
u_int32_t hash = m->m_pkthdr.pkt_flowid;
u_int32_t hindex;
u_int64_t bw;
IFCQ_LOCK_ASSERT_HELD(ifq);
/*
* If the packet doesn't have any keep state put it on the end of
* our queue. XXX this can result in out of order delivery.
*/
if (hash == 0) {
if (cl->cl_head)
b = cl->cl_head->prev;
else
b = &cl->cl_buckets[0];
} else {
hindex = (hash & cl->cl_nbucket_mask);
b = &cl->cl_buckets[hindex];
}
/*
* Add the bucket to the end of the circular list of active buckets.
*
* As a special case we add the bucket to the beginning of the list
* instead of the end if it was not previously on the list and if
* its traffic is less then the hog level.
*/
if (b->in_use == 0) {
b->in_use = 1;
if (cl->cl_head == NULL) {
cl->cl_head = b;
b->next = b;
b->prev = b;
} else {
b->next = cl->cl_head;
b->prev = cl->cl_head->prev;
b->prev->next = b;
b->next->prev = b;
if (b->bw_delta && cl->cl_hogs_m1) {
bw = b->bw_bytes * machclk_freq / b->bw_delta;
if (bw < cl->cl_hogs_m1)
cl->cl_head = b;
}
}
}
#if CLASSQ_RIO
if (cl->cl_qtype == Q_RIO)
return (rio_addq(cl->cl_rio, &b->queue, m, t));
else
#endif /* CLASSQ_RIO */
#if CLASSQ_RED
if (cl->cl_qtype == Q_RED)
return (red_addq(cl->cl_red, &b->queue, m, t));
else
#endif /* CLASSQ_RED */
#if CLASSQ_BLUE
if (cl->cl_qtype == Q_BLUE)
return (blue_addq(cl->cl_blue, &b->queue, m, t));
else
#endif /* CLASSQ_BLUE */
if (cl->cl_qtype == Q_SFB) {
if (cl->cl_sfb == NULL) {
struct ifnet *ifp = FAIRQIF_IFP(cl->cl_fif);
VERIFY(cl->cl_flags & FARF_LAZY);
IFCQ_CONVERT_LOCK(ifq);
cl->cl_sfb = sfb_alloc(ifp, cl->cl_handle,
cl->cl_qlimit, cl->cl_qflags);
if (cl->cl_sfb == NULL) {
/* fall back to droptail */
cl->cl_qtype = Q_DROPTAIL;
cl->cl_flags &= ~FARF_SFB;
cl->cl_qflags &= ~(SFBF_ECN | SFBF_FLOWCTL);
log(LOG_ERR, "%s: %s SFB lazy allocation "
"failed for qid=%d pri=%d, falling back "
"to DROPTAIL\n", if_name(ifp),
fairq_style(cl->cl_fif), cl->cl_handle,
cl->cl_pri);
}
}
if (cl->cl_sfb != NULL)
return (sfb_addq(cl->cl_sfb, &b->queue, m, t));
} else if (qlen(&b->queue) >= qlimit(&b->queue)) {
IFCQ_CONVERT_LOCK(ifq);
m_freem(m);
return (CLASSQEQ_DROPPED);
}
#if PF_ECN
if (cl->cl_flags & FARF_CLEARDSCP)
write_dsfield(m, t, 0);
#endif /* PF_ECN */
_addq(&b->queue, m);
return (0);
}
static struct priq_class *
priq_class_create(struct priq_if *pif, int pri, int qlimit, int flags, int qid)
{
struct priq_class *cl;
int s;
#ifndef ALTQ_RED
if (flags & PRCF_RED) {
#ifdef ALTQ_DEBUG
printf("priq_class_create: RED not configured for PRIQ!\n");
#endif
return (NULL);
}
#endif
if ((cl = pif->pif_classes[pri]) != NULL) {
/* modify the class instead of creating a new one */
s = splnet();
if (!qempty(cl->cl_q))
priq_purgeq(cl);
splx(s);
#ifdef ALTQ_RIO
if (q_is_rio(cl->cl_q))
rio_destroy((rio_t *)cl->cl_red);
#endif
#ifdef ALTQ_RED
if (q_is_red(cl->cl_q))
red_destroy(cl->cl_red);
#endif
} else {
cl = malloc(sizeof(struct priq_class), M_DEVBUF,
M_WAITOK|M_ZERO);
if (cl == NULL)
return (NULL);
cl->cl_q = malloc(sizeof(class_queue_t), M_DEVBUF,
M_WAITOK|M_ZERO);
if (cl->cl_q == NULL)
goto err_ret;
}
pif->pif_classes[pri] = cl;
if (flags & PRCF_DEFAULTCLASS)
pif->pif_default = cl;
if (qlimit == 0)
qlimit = 50; /* use default */
qlimit(cl->cl_q) = qlimit;
qtype(cl->cl_q) = Q_DROPTAIL;
qlen(cl->cl_q) = 0;
cl->cl_flags = flags;
cl->cl_pri = pri;
if (pri > pif->pif_maxpri)
pif->pif_maxpri = pri;
cl->cl_pif = pif;
cl->cl_handle = qid;
#ifdef ALTQ_RED
if (flags & (PRCF_RED|PRCF_RIO)) {
int red_flags, red_pkttime;
red_flags = 0;
if (flags & PRCF_ECN)
red_flags |= REDF_ECN;
#ifdef ALTQ_RIO
if (flags & PRCF_CLEARDSCP)
red_flags |= RIOF_CLEARDSCP;
#endif
if (pif->pif_bandwidth < 8)
red_pkttime = 1000 * 1000 * 1000; /* 1 sec */
else
red_pkttime = (int64_t)pif->pif_ifq->altq_ifp->if_mtu
* 1000 * 1000 * 1000 / (pif->pif_bandwidth / 8);
#ifdef ALTQ_RIO
if (flags & PRCF_RIO) {
cl->cl_red = (red_t *)rio_alloc(0, NULL,
red_flags, red_pkttime);
if (cl->cl_red != NULL)
qtype(cl->cl_q) = Q_RIO;
} else
#endif
if (flags & PRCF_RED) {
cl->cl_red = red_alloc(0, 0,
qlimit(cl->cl_q) * 10/100,
qlimit(cl->cl_q) * 30/100,
red_flags, red_pkttime);
if (cl->cl_red != NULL)
qtype(cl->cl_q) = Q_RED;
}
}
#endif /* ALTQ_RED */
return (cl);
err_ret:
if (cl->cl_red != NULL) {
#ifdef ALTQ_RIO
if (q_is_rio(cl->cl_q))
rio_destroy((rio_t *)cl->cl_red);
#endif
#ifdef ALTQ_RED
if (q_is_red(cl->cl_q))
red_destroy(cl->cl_red);
#endif
}
if (cl->cl_q != NULL)
free(cl->cl_q, M_DEVBUF);
free(cl, M_DEVBUF);
return (NULL);
}
/*
* rm_class_t *
* rmc_newclass(...) - Create a new resource management class at priority
* 'pri' on the interface given by 'ifd'.
*
* nsecPerByte is the data rate of the interface in nanoseconds/byte.
* E.g., 800 for a 10Mb/s ethernet. If the class gets less
* than 100% of the bandwidth, this number should be the
* 'effective' rate for the class. Let f be the
* bandwidth fraction allocated to this class, and let
* nsPerByte be the data rate of the output link in
* nanoseconds/byte. Then nsecPerByte is set to
* nsPerByte / f. E.g., 1600 (= 800 / .5)
* for a class that gets 50% of an ethernet's bandwidth.
*
* action the routine to call when the class is over limit.
*
* maxq max allowable queue size for class (in packets).
*
* parent parent class pointer.
*
* borrow class to borrow from (should be either 'parent' or null).
*
* maxidle max value allowed for class 'idle' time estimate (this
* parameter determines how large an initial burst of packets
* can be before overlimit action is invoked.
*
* offtime how long 'delay' action will delay when class goes over
* limit (this parameter determines the steady-state burst
* size when a class is running over its limit).
*
* Maxidle and offtime have to be computed from the following: If the
* average packet size is s, the bandwidth fraction allocated to this
* class is f, we want to allow b packet bursts, and the gain of the
* averaging filter is g (= 1 - 2^(-RM_FILTER_GAIN)), then:
*
* ptime = s * nsPerByte * (1 - f) / f
* maxidle = ptime * (1 - g^b) / g^b
* minidle = -ptime * (1 / (f - 1))
* offtime = ptime * (1 + 1/(1 - g) * (1 - g^(b - 1)) / g^(b - 1)
*
* Operationally, it's convenient to specify maxidle & offtime in units
* independent of the link bandwidth so the maxidle & offtime passed to
* this routine are the above values multiplied by 8*f/(1000*nsPerByte).
* (The constant factor is a scale factor needed to make the parameters
* integers. This scaling also means that the 'unscaled' values of
* maxidle*nsecPerByte/8 and offtime*nsecPerByte/8 will be in microseconds,
* not nanoseconds.) Also note that the 'idle' filter computation keeps
* an estimate scaled upward by 2^RM_FILTER_GAIN so the passed value of
* maxidle also must be scaled upward by this value. Thus, the passed
* values for maxidle and offtime can be computed as follows:
*
* maxidle = maxidle * 2^RM_FILTER_GAIN * 8 / (1000 * nsecPerByte)
* offtime = offtime * 8 / (1000 * nsecPerByte)
*
* When USE_HRTIME is employed, then maxidle and offtime become:
* maxidle = maxilde * (8.0 / nsecPerByte);
* offtime = offtime * (8.0 / nsecPerByte);
*/
struct rm_class *
rmc_newclass(int pri, struct rm_ifdat *ifd, u_int nsecPerByte,
void (*action)(rm_class_t *, rm_class_t *), int maxq,
struct rm_class *parent, struct rm_class *borrow, u_int maxidle,
int minidle, u_int offtime, int pktsize, int flags)
{
struct rm_class *cl;
struct rm_class *peer;
if (pri >= RM_MAXPRIO)
return (NULL);
#ifndef ALTQ_RED
if (flags & RMCF_RED) {
#ifdef ALTQ_DEBUG
kprintf("rmc_newclass: RED not configured for CBQ!\n");
#endif
return (NULL);
}
#endif
#ifndef ALTQ_RIO
if (flags & RMCF_RIO) {
#ifdef ALTQ_DEBUG
kprintf("rmc_newclass: RIO not configured for CBQ!\n");
#endif
return (NULL);
}
#endif
cl = kmalloc(sizeof(*cl), M_ALTQ, M_WAITOK | M_ZERO);
callout_init(&cl->callout_);
cl->q_ = kmalloc(sizeof(*cl->q_), M_ALTQ, M_WAITOK | M_ZERO);
/*
* Class initialization.
*/
cl->children_ = NULL;
cl->parent_ = parent;
cl->borrow_ = borrow;
cl->leaf_ = 1;
cl->ifdat_ = ifd;
cl->pri_ = pri;
cl->allotment_ = RM_NS_PER_SEC / nsecPerByte; /* Bytes per sec */
cl->depth_ = 0;
cl->qthresh_ = 0;
cl->ns_per_byte_ = nsecPerByte;
qlimit(cl->q_) = maxq;
qtype(cl->q_) = Q_DROPHEAD;
qlen(cl->q_) = 0;
cl->flags_ = flags;
#if 1 /* minidle is also scaled in ALTQ */
cl->minidle_ = (minidle * (int)nsecPerByte) / 8;
if (cl->minidle_ > 0)
cl->minidle_ = 0;
#else
cl->minidle_ = minidle;
#endif
cl->maxidle_ = (maxidle * nsecPerByte) / 8;
if (cl->maxidle_ == 0)
cl->maxidle_ = 1;
#if 1 /* offtime is also scaled in ALTQ */
cl->avgidle_ = cl->maxidle_;
cl->offtime_ = ((offtime * nsecPerByte) / 8) >> RM_FILTER_GAIN;
if (cl->offtime_ == 0)
cl->offtime_ = 1;
#else
cl->avgidle_ = 0;
cl->offtime_ = (offtime * nsecPerByte) / 8;
#endif
cl->overlimit = action;
#ifdef ALTQ_RED
if (flags & (RMCF_RED|RMCF_RIO)) {
int red_flags, red_pkttime;
red_flags = 0;
if (flags & RMCF_ECN)
red_flags |= REDF_ECN;
#ifdef ALTQ_RIO
if (flags & RMCF_CLEARDSCP)
red_flags |= RIOF_CLEARDSCP;
#endif
red_pkttime = nsecPerByte * pktsize / 1000;
if (flags & RMCF_RED) {
cl->red_ = red_alloc(0, 0,
qlimit(cl->q_) * 10/100,
qlimit(cl->q_) * 30/100,
red_flags, red_pkttime);
if (cl->red_ != NULL)
qtype(cl->q_) = Q_RED;
}
#ifdef ALTQ_RIO
else {
cl->red_ = (red_t *)rio_alloc(0, NULL,
red_flags, red_pkttime);
if (cl->red_ != NULL)
qtype(cl->q_) = Q_RIO;
}
#endif
}
#endif /* ALTQ_RED */
/*
* put the class into the class tree
*/
crit_enter();
if ((peer = ifd->active_[pri]) != NULL) {
/* find the last class at this pri */
cl->peer_ = peer;
while (peer->peer_ != ifd->active_[pri])
peer = peer->peer_;
peer->peer_ = cl;
} else {
ifd->active_[pri] = cl;
cl->peer_ = cl;
}
if (cl->parent_) {
cl->next_ = parent->children_;
parent->children_ = cl;
parent->leaf_ = 0;
}
/*
* Compute the depth of this class and its ancestors in the class
* hierarchy.
*/
rmc_depth_compute(cl);
/*
* If CBQ's WRR is enabled, then initialize the class WRR state.
*/
if (ifd->wrr_) {
ifd->num_[pri]++;
ifd->alloc_[pri] += cl->allotment_;
rmc_wrr_set_weights(ifd);
}
crit_exit();
return (cl);
}
struct hfsc_class *
hfsc_class_create(struct hfsc_if *hif, struct service_curve *rsc,
struct service_curve *fsc, struct service_curve *usc,
struct hfsc_class *parent, int qlimit, int flags, int qid)
{
struct hfsc_class *cl, *p;
int i, s;
if (hif->hif_classes >= HFSC_MAX_CLASSES)
return (NULL);
#ifndef ALTQ_RED
if (flags & HFCF_RED) {
#ifdef ALTQ_DEBUG
printf("hfsc_class_create: RED not configured for HFSC!\n");
#endif
return (NULL);
}
#endif
MALLOC(cl, struct hfsc_class *, sizeof(struct hfsc_class),
M_DEVBUF, M_WAITOK);
if (cl == NULL)
return (NULL);
bzero(cl, sizeof(struct hfsc_class));
MALLOC(cl->cl_q, class_queue_t *, sizeof(class_queue_t),
M_DEVBUF, M_WAITOK);
if (cl->cl_q == NULL)
goto err_ret;
bzero(cl->cl_q, sizeof(class_queue_t));
cl->cl_actc = actlist_alloc();
if (cl->cl_actc == NULL)
goto err_ret;
if (qlimit == 0)
qlimit = 50; /* use default */
qlimit(cl->cl_q) = qlimit;
qtype(cl->cl_q) = Q_DROPTAIL;
qlen(cl->cl_q) = 0;
cl->cl_flags = flags;
#ifdef ALTQ_RED
if (flags & (HFCF_RED|HFCF_RIO)) {
int red_flags, red_pkttime;
u_int m2;
m2 = 0;
if (rsc != NULL && rsc->m2 > m2)
m2 = rsc->m2;
if (fsc != NULL && fsc->m2 > m2)
m2 = fsc->m2;
if (usc != NULL && usc->m2 > m2)
m2 = usc->m2;
red_flags = 0;
if (flags & HFCF_ECN)
red_flags |= REDF_ECN;
#ifdef ALTQ_RIO
if (flags & HFCF_CLEARDSCP)
red_flags |= RIOF_CLEARDSCP;
#endif
if (m2 < 8)
red_pkttime = 1000 * 1000 * 1000; /* 1 sec */
else
red_pkttime = (int64_t)hif->hif_ifq->altq_ifp->if_mtu
* 1000 * 1000 * 1000 / (m2 / 8);
if (flags & HFCF_RED) {
cl->cl_red = red_alloc(0, 0,
qlimit(cl->cl_q) * 10/100,
qlimit(cl->cl_q) * 30/100,
red_flags, red_pkttime);
if (cl->cl_red != NULL)
qtype(cl->cl_q) = Q_RED;
}
#ifdef ALTQ_RIO
else {
cl->cl_red = (red_t *)rio_alloc(0, NULL,
red_flags, red_pkttime);
if (cl->cl_red != NULL)
qtype(cl->cl_q) = Q_RIO;
}
#endif
}
#endif /* ALTQ_RED */
if (rsc != NULL && (rsc->m1 != 0 || rsc->m2 != 0)) {
MALLOC(cl->cl_rsc, struct internal_sc *,
sizeof(struct internal_sc), M_DEVBUF, M_WAITOK);
if (cl->cl_rsc == NULL)
goto err_ret;
sc2isc(rsc, cl->cl_rsc);
rtsc_init(&cl->cl_deadline, cl->cl_rsc, 0, 0);
rtsc_init(&cl->cl_eligible, cl->cl_rsc, 0, 0);
}
static struct qfq_class *
qfq_class_create(struct qfq_if *qif, u_int32_t weight, u_int32_t qlimit,
u_int32_t flags, u_int32_t maxsz, u_int32_t qid, classq_pkt_type_t ptype)
{
struct ifnet *ifp;
struct ifclassq *ifq;
struct qfq_group *grp;
struct qfq_class *cl;
u_int32_t w; /* approximated weight */
int i;
IFCQ_LOCK_ASSERT_HELD(qif->qif_ifq);
if (qif->qif_classes >= qif->qif_maxclasses) {
log(LOG_ERR, "%s: %s out of classes! (max %d)\n",
if_name(QFQIF_IFP(qif)), qfq_style(qif),
qif->qif_maxclasses);
return (NULL);
}
ifq = qif->qif_ifq;
ifp = QFQIF_IFP(qif);
cl = zalloc(qfq_cl_zone);
if (cl == NULL)
return (NULL);
bzero(cl, qfq_cl_size);
if (qlimit == 0 || qlimit > IFCQ_MAXLEN(ifq)) {
qlimit = IFCQ_MAXLEN(ifq);
if (qlimit == 0)
qlimit = DEFAULT_QLIMIT; /* use default */
}
_qinit(&cl->cl_q, Q_DROPTAIL, qlimit, ptype);
cl->cl_qif = qif;
cl->cl_flags = flags;
cl->cl_handle = qid;
/*
* Find a free slot in the class table. If the slot matching
* the lower bits of qid is free, use this slot. Otherwise,
* use the first free slot.
*/
i = qid % qif->qif_maxclasses;
if (qif->qif_class_tbl[i] == NULL) {
qif->qif_class_tbl[i] = cl;
} else {
for (i = 0; i < qif->qif_maxclasses; i++) {
if (qif->qif_class_tbl[i] == NULL) {
qif->qif_class_tbl[i] = cl;
break;
}
}
if (i == qif->qif_maxclasses) {
zfree(qfq_cl_zone, cl);
return (NULL);
}
}
w = weight;
VERIFY(w > 0 && w <= QFQ_MAX_WEIGHT);
cl->cl_lmax = maxsz;
cl->cl_inv_w = (QFQ_ONE_FP / w);
w = (QFQ_ONE_FP / cl->cl_inv_w);
VERIFY(qif->qif_wsum + w <= QFQ_MAX_WSUM);
i = qfq_calc_index(cl, cl->cl_inv_w, cl->cl_lmax);
VERIFY(i <= QFQ_MAX_INDEX);
grp = qif->qif_groups[i];
if (grp == NULL) {
grp = _MALLOC(sizeof (*grp), M_DEVBUF, M_WAITOK|M_ZERO);
if (grp != NULL) {
grp->qfg_index = i;
grp->qfg_slot_shift =
QFQ_MTU_SHIFT + QFQ_FRAC_BITS - (QFQ_MAX_INDEX - i);
grp->qfg_slots = _MALLOC(sizeof (struct qfq_class *) *
qif->qif_maxslots, M_DEVBUF, M_WAITOK|M_ZERO);
if (grp->qfg_slots == NULL) {
log(LOG_ERR, "%s: %s unable to allocate group "
"slots for index %d\n", if_name(ifp),
qfq_style(qif), i);
}
} else {
log(LOG_ERR, "%s: %s unable to allocate group for "
"qid=%d\n", if_name(ifp), qfq_style(qif),
cl->cl_handle);
}
if (grp == NULL || grp->qfg_slots == NULL) {
qif->qif_class_tbl[qid % qif->qif_maxclasses] = NULL;
if (grp != NULL)
_FREE(grp, M_DEVBUF);
zfree(qfq_cl_zone, cl);
return (NULL);
} else {
qif->qif_groups[i] = grp;
}
}
cl->cl_grp = grp;
qif->qif_wsum += w;
/* XXX cl->cl_S = qif->qif_V; ? */
/* XXX compute qif->qif_i_wsum */
qif->qif_classes++;
if (flags & QFCF_DEFAULTCLASS)
qif->qif_default = cl;
if (flags & QFCF_SFB) {
cl->cl_qflags = 0;
if (flags & QFCF_ECN) {
cl->cl_qflags |= SFBF_ECN;
}
if (flags & QFCF_FLOWCTL) {
cl->cl_qflags |= SFBF_FLOWCTL;
}
if (flags & QFCF_DELAYBASED) {
cl->cl_qflags |= SFBF_DELAYBASED;
}
if (!(cl->cl_flags & QFCF_LAZY))
cl->cl_sfb = sfb_alloc(ifp, cl->cl_handle,
qlimit(&cl->cl_q), cl->cl_qflags);
if (cl->cl_sfb != NULL || (cl->cl_flags & QFCF_LAZY))
qtype(&cl->cl_q) = Q_SFB;
}
if (pktsched_verbose) {
log(LOG_DEBUG, "%s: %s created qid=%d grp=%d weight=%d "
"qlimit=%d flags=%b\n", if_name(ifp), qfq_style(qif),
cl->cl_handle, cl->cl_grp->qfg_index, weight, qlimit,
flags, QFCF_BITS);
}
return (cl);
}
static inline int
priq_addq(struct priq_class *cl, struct mbuf *m, struct pf_mtag *t)
{
struct priq_if *pif = cl->cl_pif;
struct ifclassq *ifq = pif->pif_ifq;
IFCQ_LOCK_ASSERT_HELD(ifq);
#if CLASSQ_RIO
if (q_is_rio(&cl->cl_q))
return (rio_addq(cl->cl_rio, &cl->cl_q, m, t));
else
#endif /* CLASSQ_RIO */
#if CLASSQ_RED
if (q_is_red(&cl->cl_q))
return (red_addq(cl->cl_red, &cl->cl_q, m, t));
else
#endif /* CLASSQ_RED */
#if CLASSQ_BLUE
if (q_is_blue(&cl->cl_q))
return (blue_addq(cl->cl_blue, &cl->cl_q, m, t));
else
#endif /* CLASSQ_BLUE */
if (q_is_sfb(&cl->cl_q)) {
if (cl->cl_sfb == NULL) {
struct ifnet *ifp = PRIQIF_IFP(pif);
VERIFY(cl->cl_flags & PRCF_LAZY);
cl->cl_flags &= ~PRCF_LAZY;
IFCQ_CONVERT_LOCK(ifq);
cl->cl_sfb = sfb_alloc(ifp, cl->cl_handle,
qlimit(&cl->cl_q), cl->cl_qflags);
if (cl->cl_sfb == NULL) {
/* fall back to droptail */
qtype(&cl->cl_q) = Q_DROPTAIL;
cl->cl_flags &= ~PRCF_SFB;
cl->cl_qflags &= ~(SFBF_ECN | SFBF_FLOWCTL);
log(LOG_ERR, "%s: %s SFB lazy allocation "
"failed for qid=%d pri=%d, falling back "
"to DROPTAIL\n", if_name(ifp),
priq_style(pif), cl->cl_handle,
cl->cl_pri);
} else if (pif->pif_throttle != IFNET_THROTTLE_OFF) {
/* if there's pending throttling, set it */
cqrq_throttle_t tr = { 1, pif->pif_throttle };
int err = priq_throttle(pif, &tr);
if (err == EALREADY)
err = 0;
if (err != 0) {
tr.level = IFNET_THROTTLE_OFF;
(void) priq_throttle(pif, &tr);
}
}
}
if (cl->cl_sfb != NULL)
return (sfb_addq(cl->cl_sfb, &cl->cl_q, m, t));
} else if (qlen(&cl->cl_q) >= qlimit(&cl->cl_q)) {
IFCQ_CONVERT_LOCK(ifq);
m_freem(m);
return (CLASSQEQ_DROPPED);
}
#if PF_ECN
if (cl->cl_flags & PRCF_CLEARDSCP)
write_dsfield(m, t, 0);
#endif /* PF_ECN */
_addq(&cl->cl_q, m);
return (0);
}
static struct priq_class *
priq_class_create(struct priq_if *pif, int pri, u_int32_t qlimit,
int flags, u_int32_t qid)
{
struct ifnet *ifp;
struct ifclassq *ifq;
struct priq_class *cl;
IFCQ_LOCK_ASSERT_HELD(pif->pif_ifq);
/* Sanitize flags unless internally configured */
if (pif->pif_flags & PRIQIFF_ALTQ)
flags &= PRCF_USERFLAGS;
#if !CLASSQ_RED
if (flags & PRCF_RED) {
log(LOG_ERR, "%s: %s RED not available!\n",
if_name(PRIQIF_IFP(pif)), priq_style(pif));
return (NULL);
}
#endif /* !CLASSQ_RED */
#if !CLASSQ_RIO
if (flags & PRCF_RIO) {
log(LOG_ERR, "%s: %s RIO not available!\n",
if_name(PRIQIF_IFP(pif)), priq_style(pif));
return (NULL);
}
#endif /* CLASSQ_RIO */
#if !CLASSQ_BLUE
if (flags & PRCF_BLUE) {
log(LOG_ERR, "%s: %s BLUE not available!\n",
if_name(PRIQIF_IFP(pif)), priq_style(pif));
return (NULL);
}
#endif /* CLASSQ_BLUE */
/* These are mutually exclusive */
if ((flags & (PRCF_RED|PRCF_RIO|PRCF_BLUE|PRCF_SFB)) &&
(flags & (PRCF_RED|PRCF_RIO|PRCF_BLUE|PRCF_SFB)) != PRCF_RED &&
(flags & (PRCF_RED|PRCF_RIO|PRCF_BLUE|PRCF_SFB)) != PRCF_RIO &&
(flags & (PRCF_RED|PRCF_RIO|PRCF_BLUE|PRCF_SFB)) != PRCF_BLUE &&
(flags & (PRCF_RED|PRCF_RIO|PRCF_BLUE|PRCF_SFB)) != PRCF_SFB) {
log(LOG_ERR, "%s: %s more than one RED|RIO|BLUE|SFB\n",
if_name(PRIQIF_IFP(pif)), priq_style(pif));
return (NULL);
}
ifq = pif->pif_ifq;
ifp = PRIQIF_IFP(pif);
if ((cl = pif->pif_classes[pri]) != NULL) {
/* modify the class instead of creating a new one */
if (!qempty(&cl->cl_q))
priq_purgeq(pif, cl, 0, NULL, NULL);
#if CLASSQ_RIO
if (q_is_rio(&cl->cl_q))
rio_destroy(cl->cl_rio);
#endif /* CLASSQ_RIO */
#if CLASSQ_RED
if (q_is_red(&cl->cl_q))
red_destroy(cl->cl_red);
#endif /* CLASSQ_RED */
#if CLASSQ_BLUE
if (q_is_blue(&cl->cl_q))
blue_destroy(cl->cl_blue);
#endif /* CLASSQ_BLUE */
if (q_is_sfb(&cl->cl_q) && cl->cl_sfb != NULL)
sfb_destroy(cl->cl_sfb);
cl->cl_qalg.ptr = NULL;
qtype(&cl->cl_q) = Q_DROPTAIL;
qstate(&cl->cl_q) = QS_RUNNING;
} else {
cl = zalloc(priq_cl_zone);
if (cl == NULL)
return (NULL);
bzero(cl, priq_cl_size);
}
pif->pif_classes[pri] = cl;
if (flags & PRCF_DEFAULTCLASS)
pif->pif_default = cl;
if (qlimit == 0 || qlimit > IFCQ_MAXLEN(ifq)) {
qlimit = IFCQ_MAXLEN(ifq);
if (qlimit == 0)
qlimit = DEFAULT_QLIMIT; /* use default */
}
_qinit(&cl->cl_q, Q_DROPTAIL, qlimit);
cl->cl_flags = flags;
cl->cl_pri = pri;
if (pri > pif->pif_maxpri)
pif->pif_maxpri = pri;
cl->cl_pif = pif;
cl->cl_handle = qid;
if (flags & (PRCF_RED|PRCF_RIO|PRCF_BLUE|PRCF_SFB)) {
#if CLASSQ_RED || CLASSQ_RIO
u_int64_t ifbandwidth = ifnet_output_linkrate(ifp);
int pkttime;
#endif /* CLASSQ_RED || CLASSQ_RIO */
cl->cl_qflags = 0;
if (flags & PRCF_ECN) {
if (flags & PRCF_BLUE)
cl->cl_qflags |= BLUEF_ECN;
else if (flags & PRCF_SFB)
cl->cl_qflags |= SFBF_ECN;
else if (flags & PRCF_RED)
cl->cl_qflags |= REDF_ECN;
else if (flags & PRCF_RIO)
cl->cl_qflags |= RIOF_ECN;
}
if (flags & PRCF_FLOWCTL) {
if (flags & PRCF_SFB)
cl->cl_qflags |= SFBF_FLOWCTL;
}
if (flags & PRCF_CLEARDSCP) {
if (flags & PRCF_RIO)
cl->cl_qflags |= RIOF_CLEARDSCP;
}
#if CLASSQ_RED || CLASSQ_RIO
/*
* XXX: RED & RIO should be watching link speed and MTU
* events and recompute pkttime accordingly.
*/
if (ifbandwidth < 8)
pkttime = 1000 * 1000 * 1000; /* 1 sec */
else
pkttime = (int64_t)ifp->if_mtu * 1000 * 1000 * 1000 /
(ifbandwidth / 8);
/* Test for exclusivity {RED,RIO,BLUE,SFB} was done above */
#if CLASSQ_RED
if (flags & PRCF_RED) {
cl->cl_red = red_alloc(ifp, 0, 0,
qlimit(&cl->cl_q) * 10/100,
qlimit(&cl->cl_q) * 30/100,
cl->cl_qflags, pkttime);
if (cl->cl_red != NULL)
qtype(&cl->cl_q) = Q_RED;
}
#endif /* CLASSQ_RED */
#if CLASSQ_RIO
if (flags & PRCF_RIO) {
cl->cl_rio =
rio_alloc(ifp, 0, NULL, cl->cl_qflags, pkttime);
if (cl->cl_rio != NULL)
qtype(&cl->cl_q) = Q_RIO;
}
#endif /* CLASSQ_RIO */
#endif /* CLASSQ_RED || CLASSQ_RIO */
#if CLASSQ_BLUE
if (flags & PRCF_BLUE) {
cl->cl_blue = blue_alloc(ifp, 0, 0, cl->cl_qflags);
if (cl->cl_blue != NULL)
qtype(&cl->cl_q) = Q_BLUE;
}
#endif /* CLASSQ_BLUE */
if (flags & PRCF_SFB) {
if (!(cl->cl_flags & PRCF_LAZY))
cl->cl_sfb = sfb_alloc(ifp, cl->cl_handle,
qlimit(&cl->cl_q), cl->cl_qflags);
if (cl->cl_sfb != NULL || (cl->cl_flags & PRCF_LAZY))
qtype(&cl->cl_q) = Q_SFB;
}
}
if (pktsched_verbose) {
log(LOG_DEBUG, "%s: %s created qid=%d pri=%d qlimit=%d "
"flags=%b\n", if_name(ifp), priq_style(pif),
cl->cl_handle, cl->cl_pri, qlimit, flags, PRCF_BITS);
}
return (cl);
}